decoupled-training

Proposes a decoupled training strategy that adapts normalization layers and uses precomputed features to reduce overhead in transfer learning, achieving competitive accuracy with significantly reduced training time and energy consumption.